Skip to content
← Back to Lab
Experimental

Android-Native Coding Agent

Hypothesis

A mobile-first coding agent can handle repository-level operations (branch, commit, review) from an Android device while maintaining the same quality discipline as desktop workflows.

Method

Built CodexPilot — an Android app that interfaces with local LLMs and Git operations through a structured command protocol. Tested with 3 developers across 15 real repository tasks.

Findings

Repository browsing and review work well on mobile. Full-scale code generation benefits from a hybrid approach (draft on mobile, refine on desktop). Key challenge: context window management on resource-constrained devices.

Lessons

  • Mobile review UX needs different interaction patterns than desktop — tap-to-expand diff beats scroll
  • Local model inference on Android is viable for <7B parameter models with 4-bit quantisation
  • The killer mobile feature is not writing code — it is reviewing PRs and merging with confidence
MobileDeveloper toolsAndroid